Welding is an integral part of the heavy equipment manufacturing industry, and one of the challenges technicians face during welding is porosity—the presence of cavities in the weld metal.
These cavities weaken the weld strength and require rework or even scrapping of the entire defective part.
This can be costly and time consuming. While trained weld engineers can detect porosity with auditory and visual inspection, these engineers can be difficult to find. They may also have a challenging time hearing or seeing defects in loud and smoky factory settings.
Existing automated monitoring systems often experience a high rate of false positives, slowing production and adding to costs.
John Deere, which makes agricultural, construction and forestry machinery, partnered with Intel to build an artificial intelligence (AI)-based solution for this problem.
Machine vision solutions have long been difficult to create for welding applications due to extremely harsh environments with smoke and sparks that can pose difficulties in camera placement.
To overcome these challenges, we worked with John Deere and our partners to create an AI solution with a camera placed in close range of the weld, providing insights beyond the capabilities of a human eye.
Leveraging the Intel OpenVINO (Visual Inference and Neural Network Optimization) toolkit, the solution examines the streaming video frame by frame, looking for defects.
When defects are identified by the AI model, the solution instantly switches off the welding robot so that a technician can safely intervene. Past attempts throughout the industry to deal with weld porosity issues during the welding process haven’t always been successful. If these flaws are found later in the manufacturing process, they require re-work or even scrapping of full assemblies, which can be disruptive and expensive.
Based on our pilot with Deere, the solution can detect porosity defects with up to 97.14 percent accuracy—a huge cost and productivity savings for manufacturers. The stand-alone solution has no dependency on any third-party power supply or the weld robot model, making it truly scalable.
The end-to-end, integrated system also enables manufacturers to connect new and existing welding equipment, captures multiple compute-intensive image data streams, and deploys machine learning models to edge devices.
Weld quality challenges are of course not unique to John Deere. What is unique is Deere’s approach. Engineers at the company are leaning into AI and machine vision to automate quality inspection, enabling them to detect issues as they happen to drive fast decision-making on the factory floor, automate the QA process to improve quality, reduce costs and increase factory throughput.
While working with Deere, our focus was to ensure the longevity of the solution to support other quality or efficiency needs that the company might have in the future.
This is a big takeaway for companies looking to add intelligence to existing manufacturing equipment, processes and management. It’s vital to partner with solution providers who can help solve today’s challenges while ensuring interoperability and scalability for flexibility and ease of use for the future.
“Welding is complicated. This AI solution has the potential to help us produce our ... machines more efficiently than before,” said Andy Benko, quality director in Deere’s construction & forestry unit.